[llvm-branch-commits] [llvm] release/23.x: Revert "[LV] Allow scalable epilogue VFs matching the MainLoop VF (#208686) (PR #211545)

[David Sherwood via llvm-branch-commits]

david-arm wrote:

> Reverting for the release sounds good. Should we also revert on `main` if it is a notable regression?

I personally think that's probably a good idea for now. What do you think @MacDue? Although PR #208686 is perfectly sensible, it does highlight some issues when using scalable vectors in vector epilogue loops pre-LTO when we don't know the trip count. This is the issue with bwaves, since we only discover the trip count (5) at the LTO stage. When using fixed-width vector epilogues we end up unrolling/deleting the vector epilogue and scalar loops leaving clean, straight-line code. When using scalable vectors the loops cannot be eliminated due to not knowing the value of vscale at compile time.

https://github.com/llvm/llvm-project/pull/211545